forecasting-tools 0.1.4

AI forecasting and research tools to help humans reason about and forecast the future

Last Update: Nov 3 2024

Quick Install

Install this package with pip install forecasting-tools

Overview

Demo website: https://mokoresearch.streamlit.app/

This repository contains forecasting and research tools built with Python and Streamlit. The project aims to assist users in making predictions, conducting research, and analyzing data related to hard to answer questions (especially those from Metaculus).

Key features: Tools most likely to be useful to you

• 🔍 Perplexity++ Smart Searcher: Smart Searcher for AI-powered internet powered by Exa.ai that is configurable, more accurate, able to decide on filters, able to link to exact paragraphs, and generally better than Perplexity.ai.
• 🔑 Key Factor Analysis: Key Factors Analysis for scoring, ranking, and prioritizing important variables in forecasting questions
• 🎯 Forecasting Bot: General Forecaster (Forecast Team) that integrates with the Metaculus AI benchmarking competition and really any other question you have.

Here are some other cool components and features of the project:

• Base Rate Researcher: for calculating event probabilities
• Niche List Researcher: for analyzing very specific lists of past events or items
• Fermi Estimator: for breaking down numerical estimates
• Team Manager: for automating forecasts across multiple questions
• Metaculus API Wrapper: for interacting with questions and tournaments
• AI Model Wrappers: (GPT-4, Claude 3.5, Perplexity) with reliability upgrades, cost management, structured outputs, etc
• Monetary Cost Manager: for tracking AI and API expenses

Join the discord for updates and to give feedback (btw feedback is very appreciated, even just a quick 'thank you' or 'I decided not to use the tool for reason X' is helpful to know)

Note: AI inaccuracies are expected (with some tools morethan others), and the package api isn't guaranteed to stay consistent (though I will shoot to keep it stable).

Examples

Smart Searcher

The Smart Searcher acts like an LLM with internet access. It works a lot like Perplexity.ai API, except:

• It has clickable citations that highlights and links directly to the paragraph cited using text fragments
• You can ask the AI to use filters for domain, date, and keywords
• There are options for structured output (Pydantic objects, lists, dict, list[dict], etc.)
• Concurrent search execution for faster results
• Optional detailed works cited list

from forecasting_tools import SmartSearcher

searcher = SmartSearcher(
    temperature=0,
    num_searches_to_run=2,
    num_sites_per_search=10,  # Results returned per search
    include_works_cited_list=False  # Add detailed citations at the end
)

response = await searcher.invoke(
    "What is the recent news for Apple?"
)

print(response)

Example output:

Recent news about Apple includes several significant developments:

1. Expansion in India: Apple is planning to open four more stores in India, with two in Delhi and Mumbai, and two in Bengaluru and Pune. This decision follows record revenues in India for the September 2024 quarter, driven by strong iPhone sales. Tim Cook, Apple's CEO, highlighted the enthusiasm and growth in the Indian market during the company's earnings call [1][4][5][8].

2. Product Launches: Apple is set to launch new iMac, Mac mini, and MacBook Pro models with M4 series chips on November 8, 2024. Additionally, the Vision Pro headset will be available in South Korea and the United Arab Emirates starting November 15, 2024. The second season of the Apple TV+ sci-fi series "Silo" will also premiere on November 15, 2024 [2][12].

... etc ...

You can also use structured outputs by providing a Pydantic model (or any other simpler type hint) and using the schema formatting helper:

from pydantic import BaseModel, Field
from forecasting_tools import SmartSearcher

class Company(BaseModel):
    name: str = Field(description="Full company name")
    market_cap: float = Field(description="Market capitalization in billions USD")
    key_products: list[str] = Field(description="Main products or services")
    relevance: str = Field(description="Why this company is relevant to the search")

searcher = SmartSearcher(temperature=0, num_searches_to_run=4, num_sites_per_search=10)

schema_instructions = searcher.get_schema_format_instructions_for_pydantic_type(Company)
prompt = f"""Find companies that are leading the development of autonomous vehicles.
Return as a list of companies with their details. Remember to give me a list of the schema provided.

{schema_instructions}"""

companies = await searcher.invoke_and_return_verified_type(prompt, list[Company])

for company in companies:
    print(f"\n{company.name} (${company.market_cap}B)")
    print(f"Relevance: {company.relevance}")
    print("Key Products:")
    for product in company.key_products:
        print(f"- {product}")

The schema instructions will format the Pydantic model into clear instructions for the AI about the expected output format and field descriptions.

Key Factors Researcher

The Key Factors Researcher helps identify and analyze key factors that should be considered for a forecasting question. As of last update, this is the most reliable of the tools, and gives something useful and accurate most every time. It asks a lot of questions, turns search results into a long list of bullet points, rates each bullet point on ~8 criteria, and returns the top results.

from forecasting_tools import KeyFactorsResearcher, BinaryQuestion, QuestionState

# Consider using MetaculusApi.get_question_by_id or MetaculusApi.get_question_by_url instead
question = BinaryQuestion(
    question_text="Will YouTube be blocked in Russia?",
    background_info="...", # Or 'None'
    resolution_criteria="...", # Or 'None'
    fine_print="...", # Or 'None'
    question_id=0, # The ID and state only matters if using Metaculus API calls
    question_state=QuestionState.OPEN
)

# Find key factors
key_factors = await KeyFactorsResearcher.find_key_factors(
    metaculus_question=question,
    num_key_factors_to_return=5,  # Number of final factors to return
    num_questions_to_research_with=26  # Number of research questions to generate
)

print(ScoredKeyFactor.turn_key_factors_into_markdown_list(key_factors))

Example output:

• The Russian authorities have slowed YouTube speeds to near unusable levels, indicating a potential groundwork for a future ban. Source Published on 2024-09-12
• Russian lawmaker Alexander Khinshtein stated that YouTube speeds would be deliberately slowed by up to 70% due to Google's non-compliance with Russian demands, indicating escalating measures against YouTube. Source Published on 2024-07-25
• The press secretary of President Vladimir Putin, Dmitry Peskov, denied that the authorities intended to block YouTube, attributing access issues to outdated equipment due to sanctions. Source Published on 2024-08-17
• YouTube is currently the last Western social media platform still operational in Russia, with over 93 million users in the country. Source Published on 2024-07-26
• Russian users reported mass YouTube outages amid growing official criticism, with reports of thousands of glitches in August 2024. Source Published on 2024-08-09

The simplified pydantic structure of the scored key factors is:

class ScoredKeyFactor():
    text: str
    factor_type: KeyFactorType (Pro, Con, or Base_Rate)
    citation: str
    source_publish_date: datetime | None
    url: str
    score_card: ScoreCard
    score: int
    display_text: str

Base Rate Researcher

The Base Rate Researcher helps calculate historical base rates for events. As of last update, it gives decent results around 50% of the time. It orchestrates the Niche List Researcher and the Fermi Estimator to find base rate.

from forecasting_tools import BaseRateResearcher

# Initialize researcher
researcher = BaseRateResearcher(
    "How often has Apple been successfully sued for patent violations?"
)

# Get base rate analysis
report = await researcher.make_base_rate_report()

print(f"Historical rate: {report.historical_rate:.2%}")
print(report.markdown_report)

Niche List Researcher

The Niche List Researcher helps analyze specific lists of events or items. The researcher will:

1. Generate a comprehensive list of potential matches
2. Remove duplicates
3. Fact check each item against multiple criteria
4. Return only validated items (unless include_incorrect_items=True)

from forecasting_tools import NicheListResearcher

researcher = NicheListResearcher(
    type_of_thing_to_generate="Times Apple was successfully sued for patent violations between 2000-2024"
)

fact_checked_items = await researcher.research_niche_reference_class(
    return_invalid_items=False
)

for item in fact_checked_items:
    print(item)

The simplified pydantic structure of the fact checked items is:

class FactCheckedItem():
    item_name: str
    description: str
    is_uncertain: bool | None = None
    initial_citations: list[str] | None = None
    fact_check: FactCheck
    type_description: str
    is_valid: bool
    supporting_urls: list[str]
    one_line_fact_check_summary: str

class FactCheck(BaseModel):
    criteria_assessments: list[CriteriaAssessment]
    is_valid: bool

class CriteriaAssessment():
    short_name: str
    description: str
    validity_assessment: str
    is_valid_or_unknown: bool | None
    citation_proving_assessment: str | None
    url_proving_assessment: str | None:

Fermi Estimator

The Fermi Estimator helps break down numerical estimates using Fermi estimation techniques.

from forecasting_tools import Estimator

estimator = Estimator(
    type_of_thing_to_estimate="books published worldwide each year",
    previous_research=None  # Optional: Pass in existing research
)

size, explanation = await estimator.estimate_size()

print(f"Estimate: {size:,}")
print(explanation)

Example output (Fake data with links not added):

I estimate that there are 2,750,000 'books published worldwide each year'.

Facts:

• Traditional publishers release approximately 500,000 new titles annually in English-speaking countries [1]
• China publishes around 450,000 new books annually [2]
• The global book market was valued at $92.68 billion in 2023 [3]
• Self-published titles have grown by 264% in the last 5 years [4]
• Non-English language markets account for about 50% of global publishing [5]

Estimation Steps and Assumptions:

1. Start with traditional English publishing: 500,000 titles
2. Add Chinese market: 500,000 + 450,000 = 950,000
3. Account for other major languages (50% of market): 950,000 * 2 = 1,900,000
4. Add self-published titles (estimated 45% of total): 1,900,000 * 1.45 = 2,755,000

Background Research: [Additional research details...]

Forecast Team (General Forecaster)

The Forecast Team combines multiple forecasting approaches to analyze and make predictions for Metaculus questions:

from forecasting_tools import ForecastTeam, BinaryQuestion

# Consider using MetaculusApi.get_question_by_id or MetaculusApi.get_question_by_url instead
question = BinaryQuestion(
    question_text="Will humanity be extinct by 2100?",
    background_info=None,
    resolution_criteria=None,
    fine_print=None,
    question_id=0
)

team = ForecastTeam(question=question)

report = await team.run_forecast()

print(f"Prediction: {report.prediction:.1%}")
print(report.explanation)

Team Manager

The Team Manager helps automate forecasting across multiple questions, particularly useful for Metaculus AI benchmarking and tournament participation. It can run forecasts on all open questions in a tournament, publish predictions, and benchmark forecast performance.

Note that (as of last update) benchmarking is in beta. It is helpful to have the front end to view results. There may be bias from struggles with multiple choice question wording. Benchmarks are run by comparing the community forecast to the team forecast on a semi-random filtered set of binary open questions. This isn't a perfect way to measure, but better than waiting till the end of a tournament or doing it manually.

from forecasting_tools import TeamManager

manager = TeamManager(time_to_wait_between_questions=60)

# Run and publish forecasts for all open questions in a tournament
reports = await manager.run_and_publish_forecasts_on_all_open_questions(
    tournament_id=3672
)

# Or run forecasts without publishing
reports = await manager.run_forecasts_on_all_open_questions(
    tournament_id=3672
)

# Sanity check changes by benchmarking forecast performance
# Depths: "shallow" (10 questions), "medium" (20), "deep" (30)
score = await manager.benchmark_forecast_team(
    evaluation_depth="medium"
)
print(f"Average deviation score: {score:.3f}")

Metaculus API

The Metaculus API wrapper helps interact with Metaculus questions and tournaments. Grabbing questions returns a pydantic object, and supports important information for Binary, Multiple Choice, Numeric,and Date questions.

from forecasting_tools import MetaculusApi

question = MetaculusApi.get_question_by_id(11245)  # US 2024 Election
question = MetaculusApi.get_question_by_url("https://www.metaculus.com/questions/11245/...")
questions = MetaculusApi.get_all_questions_from_tournament(
    tournament_id=3672,  # Q4 2024 Quarterly Cup
    filter_by_open=True  # Only return open questions
)
MetaculusApi.post_binary_question_prediction(
    question_id=11245,
    prediction_in_decimal=0.75  # Must be between 0.01 and 0.99
)
MetaculusApi.post_question_comment(
    question_id=11245,
    comment_text="Here's my reasoning..."
)
benchmark_questions = MetaculusApi.get_benchmark_questions(
    num_of_questions_to_return=20
) # See mention of benchmark questions in Team Manager section

AI Models

Wrapper classes on a number of AI models are provided that include, structured output, retry, cost management, token limiting, rate limiting, timeouts, and more. See forecasting_tools/ai_models/README.md for more details. As of last update, token and rate limiting are set to the publisher's limits, and needs to be extracted into an environment setting.

Models included are:

• GPT-4o
• GPT-o1
• GPT-4o-vision
• Claude 3.5 Sonnet
• Perplexity
• Metaculus Proxy (specifically for GPT-4o)

Monetary Cost Manager

The Monetary Cost Manager helps to track AI and API costs. It tracks expenses and errors if it goes over the limit. Leave the limit empty to disable the limit. It shouldn't be trusted as an exact expense, but a good estimate of costs. See forecasting_tools/ai_models/README.md for more details, and some flaws it has.

from forecasting_tools import MonetaryCostManager

max_cost = 5.00

with MonetaryCostManager(max_cost) as cost_manager:
    results = await any_tool_that_uses_ai()
    current_cost = cost_manager.current_usage
    print(f"Current cost: ${current_cost:.2f}")

Local Development

Environment Variables

The environment variables you need can be found in .env.template. Copy this template as .env and fill it in. As of last update, you only strictly need OPENAI_API_KEY and EXA_API_KEY.

Docker Dev Container

Dev containers are reliable ways to make sure environments work on everyone's machine the first try and so you don't have to spend hours setting up your environment (especially if you have docker already installed). If you would rather just use poetry, without the dev container, you can skip to "Alternatives to Docker". Otherwise, to get your development environment up and running, you need to have Docker Engine installed and running. Once you do, you can use the VSCode dev container pop-up to automatically set up everything for you.

Install Docker

For Windows and Mac, you will download Docker Desktop. For Linux, you will download Docker Engine. (NOTE: These instructions might be outdated).

First download and setup Docker Engine using the instructions at the link below for your OS:

• Windows: windows-install
• Mac: mac-install
• Linux: install
  • Do not install Docker Desktop for Linux, rather, select your Linux distribution on the left sidebar and follow the distribution specific instructions for Docker engine. Docker Desktop runs with a different environment in Linux.
  • Remember to follow the post-installation steps for Linux: linux-postinstall

Starting the container

Once Docker is installed, when you open up the project folder in VSCode, you will see a pop up noting that you have a setup for a dev container, and asking if you would like to open the folder in a container. You will want to click "open in container". This will automatically set up everything you need and bring you into the container. If the docker process times out in the middle of installing python packages you can run the .devcontiner/postinstall.sh manually. You may need to have the VSCode Docker extension and/or devcontainer extension downloaded in order for the pop up to appear.

Once you are in the container, poetry should have already installed a virtual environment. For VSCode features to use this environment, you will need to select the correct python interpreter. You can do this by pressing Ctrl + Shift + P and then typing Python: Select Interpreter. Then select the interpreter that starts with .venv.

A number of vscode extensions are installed automatically (e.g. linting). You may need to wait a little while and then reload the window after all of these extensions are installed. You can install personal vscode extensions in the dev environment.

Managing Docker

There are many ways to manager Docker containers, but generally if you download the vscode docker extension, you will be able to stop/start/remove all containers and images.

Alternatives to Docker

If you choose not to run docker, you can use poetry to set up a local virtual environment. If you are on Ubuntu, you should be able to just read through and then run .devcontainer/postinstall.sh. If you aren't on Ubuntu, check out the links in the postinstall file for where install instructions for dependencies were originally found. You may also want to take a look at VSCode extensions that would be installed (see the list in the .devcontainer/devcontainer.json file) so that some VSCode workplace settings work out of the box (e.g. automatic Black Formatting).

Running the Front End

You can run any front end folder in the front_end directory by executing streamlit run front_end/Home.py. This will start a development server for you that you can run.

Testing

This repository uses pytest and pytest-xdist. xdist spreads out all the tests between multiple threads that are each run on a separate CPU in order to speed up execution. Currently its setup to create a thread per CPU. Configuration for this is in pytest.ini. The tests are gathered afresh from each thread, so any initialization done in imports, globals, or class variables are done for each thread.

Tests are subdivided into folders based on cost and reliability, so do not run pytest without specifying which folder you want or else you will incur some large expenses from OpenAI. The more expensive tests run the tools all the way through, and on many example cases, and are not expected to get 100% success rates.